A small VAT discrepancy. A laptop on my knees. And an hour and a half to two hours that changed how I saw the machine I had worked with for most of my adult life.
I had received a letter from the Danish tax authorities about a small discrepancy in a VAT return. As I remember it, the amount was only a few hundred kroner. A minor administrative loose end, but one that still needed sorting out.
What began with that small discrepancy grew into a broader review and corrections to previous VAT returns. There were invoices to find, figures to compare and explanations to prepare. Several systems, each holding part of the answer.
I uploaded the letter to ChatGPT.
I had done that with documents before. This time I was using a desktop app: a program installed on my computer. With the relevant tools enabled and access granted, the assistant could work with my files and operate the websites and applications involved in the task.
Installing the app alone does not give it access to everything. In my case, I logged into the relevant services myself, granted access and set it to work.
There is a practical detail worth knowing: ChatGPT Work and Codex are included in ChatGPT plans, with usage limits that depend on the plan. You may already have access through a subscription you pay for. The available features and allowances still need checking, but this need not begin with buying a separate software licence.
Then I sat back with my screen split in two and watched.
Something was happening on my screen
ChatGPT worked through information in the tax portal, my accounting software and my online bank. With relatively little input from me, it gathered material, prepared VAT calculations and worked on corrections.
My information. My accounts. My real, mildly irritating problem, being dealt with in front of my eyes.
Along the way, ChatGPT suggested checking whether everything had been included, since we were submitting corrected VAT returns anyway.
That struck me as a fairly sensible suggestion.
So I asked it to go further. Review bank statements going back three or four years. Look for invoices in my email, including attachments from suppliers. Check recurring subscriptions in the accounting system: was there actually an invoice for every month I had paid? And run some spot checks to see how up to date the accounts were.
The task grew while I sat there. ChatGPT reviewed entries and made corrections in the accounting software as I followed its work on screen.
After roughly an hour and a half to two hours, the corrections and amended VAT returns had been submitted, along with an explanation to the tax authorities. There were submission receipts.
The authorities still had their processing to do, of course. Even a very industrious AI is unlikely to put the government calendar into the fast lane.
But my side of the work had reached submission. My own estimate was that the process could otherwise easily have taken a capable accountant a working day or more. That is an estimate, not a stopwatch study of the accounting profession.
I was completely blown away.
I had been the connection all along
I have spent years starting businesses, working with technology and finding new ways of doing things. I am used to software that promises to save me time.
Some of it has mostly saved time in its own sales presentation.
Because even with good systems, I was usually the one who had to connect them. I read the letter. Found the invoice. Opened the bank. Searched my email. Compared the information. Moved the figures and wrote the explanation.
The computer held the material and the tools. I had to supply the connections.
Now I could describe what I wanted investigated and done, and let a digital assistant work through several of the necessary steps. I still had decisions to make and remained responsible. But my role in the work had shifted.
Was I being brave or just an idiot?
That thought genuinely crossed my mind while I was sitting there.
I had given an AI access to private information, email and financial systems. I could watch the work, and I had logged in myself. It felt manageable. I also reasoned that a bank transfer would need my approval.
But that feeling is not a security assessment. Logging into Chrome myself does not mean all the information stays on my computer. And an additional approval for a bank transfer does not protect against every other mistake: incorrect changes, ill-judged submissions or information ending up in the wrong place.
So my experience cannot prove that it is safe to give an AI access to everything. What I can say is that I took the leap, followed the work and saw enough value to want to explore the possibilities much more seriously.
The excitement and the doubt were there together. In truth, they still are.
Then I thought about my insurance
Once I had seen this, the ideas started queuing up.
The first thing I wanted to investigate was my insurance. What am I actually paying? What does it cover? How have the premiums changed? And how does it compare with the alternatives?
I imagined giving the assistant access to policies, invoices and my history. Having it review the material, research competitors and compare the terms. Then prepare a letter to my insurer if there was a case for asking for a better price.
Coverage, excesses and exclusions would have to be part of the comparison, naturally. Finding cheap insurance is easy enough if you do not care what happens when you need it.
I am not claiming that this entire task was therefore completed. It was the first of the many possibilities I immediately wanted to try.
Then came subscriptions. Old agreements. Documents I had forgotten. Information scattered across emails, folders and websites. Work on my own websites.
Suddenly, I could see hundreds of uses.
The caretaker of my own digital archive
Over the years, we have accumulated enormous amounts of material. Contracts, correspondence, bills, ideas, research and experience. I certainly have.
And I have spent a fair amount of time looking after it all. Saving. Naming. Moving. Searching. Considering a slightly better filing system that I would definitely stick to this time.
Being the caretaker of your own digital archive can become quite a substantial second job.
The accounting experience showed me what can happen when an assistant can find and connect some of that material around a specific task. Old information becomes useful in a new way.
Which agreements have become more expensive? What did we agree with the supplier? Is the evidence already sitting in an email? What do I need to establish before I reply?
Of course, I could investigate these things before. But plenty of tasks never got done because they took too much time. Now more of them could become practical to carry out.
The summer it became real for me
If names such as Codex, ChatGPT Work and Claude Cowork mean little to you, the practical idea is easier to grasp: an AI assistant with tools to carry out work. In my case, I used a desktop app on my laptop and watched it work across the systems involved.
These products differ, and access and capabilities vary. Some tasks use direct connections to services; others involve the assistant operating applications and websites. The desktop app was how I experienced this change, rather than a requirement for every kind of AI-assisted work.
This story is about my encounter with that way of working. The summer of 2026 was when it became real for me. I am not claiming the technology was first invented or launched that summer.
I sat with my own laptop and discovered that something familiar had changed character. Like a chrysalis opening to reveal something you had not quite understood was inside.
I do not know how quickly others will experience the same thing. Nor do I know how much of my work I will eventually delegate to assistants like these.
But I know that the way I saw a computer changed that day.
And that my first thought, once the astonishment had settled, was:
What else have I been spending my time on?